Research on a hierarchical feature-based contour extraction method for spatial complex truss-like structures in aerial images

Spatial truss-like structures are three-dimensional frame structures consisting of a series of linear members and nodes, which are widely used in the power engineering. UAVs and various land-air robots are widely used in power inspection to reduce the reliance on manual labor. Extracting the contours of such objects (represented in the form of points and line segments) in aerial images will further improve the efficiency of power inspections. Because existing methods struggle to deal with the complex background interference in aerial images, an innovative contour extraction method based on hierarchical features is proposed in this paper, which uses two artificial neural networks to extract image-wise features and patch-wise features respectively and link them with patch division. The IoU of the method can reach 83.4%, which is improved by 5.4% and 50.8% compared to the lightweight semantic segmentation network and HED-based method, respectively. Meanwhile the Polygons and line segments measurement (PoLis) outperforms the other four types of methods compared with it by at least 39.1%, which overcomes the drawbacks of the other methods that have poor contour extraction ability in complex backgrounds and aerial images with numerous feature. Meanwhile, the running time of this method is 85 ms. The proposed method overcomes the shortcomings of existing contour extraction methods, improves the effect of contour extraction in complex backgrounds in power inspection scenarios, facilitating the improvement of the intelligence level of inspection robots, which in turn promotes the automation level in power engineering.